Tilburg University Robust Control Charts for Time Series ...
Transcript of Tilburg University Robust Control Charts for Time Series ...
Tilburg University
Robust Control Charts for Time Series Data
Croux, C.; Gelper, S.; Mahieu, K.
Publication date:2010
Link to publication in Tilburg University Research Portal
Citation for published version (APA):Croux, C., Gelper, S., & Mahieu, K. (2010). Robust Control Charts for Time Series Data. (CentER DiscussionPaper; Vol. 2010-107). Econometrics.
General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal
Take down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.
Download date: 13. Dec. 2021
ROBUST CONT
By
ISSN 0924
No. 2010–107
ROBUST CONTROL CHARTS FOR TIME SERIES DATA
By Christophe Croux, Sarah Gelper and Koen Mahieu
October 2010
ISSN 0924-7815
ROL CHARTS FOR TIME SERIES DATA
Gelper and Koen Mahieu
Robust control charts for time series data
Christophe Croux
K.U. Leuven & Tilburg University
Sarah Gelper
Erasmus University Rotterdam
Koen Mahieu
K.U. Leuven
Abstract
This article presents a control chart for time series data, based on the one-step-
ahead forecast errors of the Holt-Winters forecasting method. We use robust
techniques to prevent that outliers affect the estimation of the control limits
of the chart. Moreover, robustness is important to maintain the reliability of
the control chart after the occurrence of alarm observations. The properties
of the new control chart are examined in a simulation study and on a real
data example.
Keywords: Control chart, Holt-Winters, Non-stationary time series, Out-
lier detection, Robustness, Statistical process control.
1 Introduction
Control charts are used to detect anomalies in processes. They are most often
used to monitor production-related processes. Samples taken from such processes
are assumed to be independent and identically distributed. Many business-related
processes, for instance sales volumes or product prices, behave very differently. They
typically contain a trend, local or global, and serial correlation. In this paper we
propose a control chart aimed at detecting aberrant observations, or outliers, in
such time series. The control chart needs to be robust, meaning that the presence
of outliers in the series should not harm its performance. Outliers may be present
in the “training period”, being the part of the series used to determine the control
limits, or in the “test period”, being the part of the series where outliers should be
flagged as alarm observations.
1
The control chart we propose belongs to the class of special-cause control charts
(Alwan and Roberts, 1988), where forecast errors of a prediction method are subject
to regular control chart techniques. If an observation has a large deviation from its
predicted value, an unexpected event occurred, and an alarm should be given. A
large difference between the observed and predicted value implies an unusually large
forecasting error, and the corresponding observation will be flagged as an outlier in
the control chart of the forecast errors. In this paper, we apply the Holt-Winters
forecast method, which is a widely used and simple procedure to forecast time series
containing trends and serial correlation. Moreover, it is known to yield good forecast
performance for many types of times series encountered in business and industry
(see Gardner (2006) and De Gooijer and Hyndman (2006) for recent applications).
Every single forecast error is then monitored on an X-chart for individual data.
Since the control chart should be resistant to outliers, a robust version of both the
Holt-Winters forecasting procedure and the X-chart needs to be used.
Several types of control charts for time series data have been proposed in the
literature. To deal with serial correlation in stationary processes ARMA-charts
(Jiang, Tsui, and Woodall, 2000) were proposed, or data transformations (Wang,
2005). CuScore charts (Box and Ramırez, 1992) can be of use for non-stationary
series; see Nembhard and Changpetch (2007) for a recent application. Vander Wiel
(1996) proposed control charts for processes that wander, based on integrated mov-
ing average models. The Holt-Winters method is an important example of this
approach. As the other proposals, the approach of Vander Wiel (1996) is sensitive
to outliers in the training sample and test sample. This problem is remediated by
following the robust approach taken in this paper.
Robust versions of the standard X and R chart are given in Rocke (1989), Rocke
(1992) and Tatum (1997), where the mean and scale of the process are estimated ro-
bustly for setting up the control limits. More recently, robust control charts for mul-
tivariate processes were proposed, see Vargas (2003), Alfaro and Ortega (2009), and
Stefatos and Hamza (2009). This paper contributes to this literature by providing a
robust control chart for time series typically encountered in business environments,
containing stochastic trends, outliers, and strong serial correlation.
In Section 2 control charts based on the Holt-Winters method are briefly ex-
plained, while Section 3 describes the robust version. The Holt-Winters method,
both the standard and the robust version, has the advantage of being easy to imple-
ment by means of a simple updating scheme. Section 4 presents a simulation study
that demonstrates the advantages of a robust approach. First, the robust version
2
gives reliable control limits in cases where the training sample contains outliers.
Secondly, if an isolated outlier occurs in the test sample, the non-robust control
chart might yield a sequence of false alarms right after the occurrence of the outlier.
The robust approach does not suffer from this drawback. Section 5 presents an
application with real data. Finally, some conclusions are made in Section 6.
2 Control charts using the Holt-Winters method
In this section, we construct a control chart for a time series {yt}. At every time
point we make a one-step-ahead forecast by the Holt-Winters forecasting algorithm.
The one-step-ahead forecast errors of this procedure are plotted on a control chart.
Control limits to monitor new incoming observations are constructed from the train-
ing sample.
2.1 The forecasting algorithm
Consider a time series observed up to time point t − 1. The Holt-Winters method
makes a prediction of the series at time t, denoted by yt|t−1. Once the true value yt
is observed, the one-step-ahead forecast error is given by
et = yt − yt|t−1. (1)
Denote the local level of the series by {αt} and the local trend by {βt}. The
Holt-Winters algorithm estimates these unknown values by
αt = λ1 yt + (1 − λ1) (αt−1 + βt−1)
βt = λ2 (αt − αt−1) + (1 − λ2) βt−1
(2)
resulting in the forecast
yt|t−1 = αt−1 + βt−1 . (3)
The parameters λ1 and λ2 in (2) are smoothing parameters, with values between
zero and one. The larger the smoothing parameters, the less the local level and/or
local trend series are smoothed, and the more weight the current value of yt has on
the prediction of the next value.
The recursive relations (2) are started up after a short startup period of length
m. A linear trend line is fitted to the data in the startup period for estimating the
3
local level and trend at t = m. The smoothing parameters λ1 and λ2 are selected
by minimizing the sum of squared forecast errors over a training period of length n:
(λopt
1 , λopt
2 ) = argmin(λ1,λ2)
n∑
t=m+1
(
yt − yt|t−1
)2. (4)
Note that the observations in the startup period are not used for determining the
smoothing parameters.
2.2 Control chart
As motivated in Section 1, the one-step-ahead forecast errors et = yt − yt|t−1 are
monitored to detect anomalies in the outcome of the process. We proceed as for a
regular X-chart, and assume normality of the forecast errors. The control limits are
computed from the forecast errors et in the training period, with t = m + 1, . . . , n.
Denote S2 the mean of the squared prediction errors in the training period, excluding
the startup values:
S2 =
∑nt=m+1 e
2t
n−m.
Since the target value for the forecast errors is zero, the following (1−α) Upper and
Lower Control Limits are obtained:
UCL = zα/2 × S
LCL = − zα/2 × S ,(5)
with zα/2 the (1 − α/2)-quantile of the standard normal distribution.
For every new observations yt in the test period, where t > n, one updates the
recursion relations (2) and computes the new prediction error et as in (1). The
value of et is then plotted in the control chart. If it falls out of the control limits,
then the observed value deviates substantially from the predicted value, indicating
an unexpected change in the process. Note that the observations in the test period
influence future prediction errors via the Holt-Winters forecast formulas in (2), but
do not alter the control limits.
It is of crucial importance that the process is completely in control during the
training period, otherwise the scale estimate S is inflated by the presence of outliers
in the training period. If outliers are present in the training sample, it is necessary
to follow a robust approach, as discussed in the next section.
4
3 Control charts using the Robust Holt-Winters
method
The control chart presented in the previous section is sensitive to outlying obser-
vations in several ways. First, the Holt-Winters forecasts are adversely affected by
outliers, since the estimated local level and trend in (2) depend on past values of the
observed series, and thus also on possible outliers. Second, if outliers are present in
the training sample, the selected smoothing parameters λ1 and λ2 might be biased.
Finally, due to an inflated scale of the forecast errors, outliers may lead to wider
control limits.
3.1 The forecasting algorithm
The shortcomings of the standard Holt-Winters algorithm have been addressed by
Gelper et al. (2009). They present a robust version of Holt-Winters method, consist-
ing of the application of the standard procedure on a cleaned version of the observed
series. Furthermore, they stress the importance of a robust selection of the smooth-
ing parameters and a robust estimation of the starting values of the algorithm. We
briefly review their approach.
A cleaned version y∗t of the time series yt is obtained as
y∗t = ψk
(
yt − αt−1 − βt−1
σt
)
σt + αt−1 + βt−1 (6)
where ψk(y) = max(−k,min(y, k)) is the Huber ψ-function with boundary value
k, and σt is a local scale estimate of the one-step-ahead forecast errors. We set
k = 2, such that the influence of forecast errors larger than two times the local
scale estimate is bounded. Note that for k tending to infinity, one gets the standard
Holt-Winters method again. The local scale estimate is computed recursively as
σ2t = λσρk
(
yt − αt−1 − βt−1
σt−1
)
σ2t−1 + (1 − λσ)σ2
t−1 (7)
where ρk(·) is a bounded loss function with boundary value k = 2, see Gelper et al.
(2009). Forecast errors larger than two times the local scale estimate σt−1 will have
a bounded influence on the next scale estimate. We allow for a slowly varying scale,
which is translated in a choice of λσ rather close to zero, for example λσ = 0.3.
5
The recursive scheme of the robust Holt-Winters algorithm then becomes, anal-
ogous to (2),
αt = λ1 y∗t + (1 − λ1)(αt−1 + βt−1)
βt = λ2 (αt − αt−1) + (1 − λ2) βt−1 ,(8)
where y∗t follows equation (6).
There are two further issues that need to be addressed. First, the startup values
αm and βm need to be computed using a robust regression fit to the data in the
startup period. For this robust fit the repeated median estimators is used. The
median absolute deviation of the residuals with respect to this robust fit yields the
starting value σ2m for the scale estimates. The second issue is the robust selection of
the smoothing parameters. Instead of minimizing the sum of squares of the one-step-
ahead forecast errors over the training sample, we minimize a robust performance
measure
(λopt
1 , λopt
2 ) = argmin(λ1,λ2)
s20
n∑
t=m+1
ρk
(
yt − yt|t−1
s0
)
. (9)
This performance measure corresponds to a τ -scale estimator, see for example
Maronna et al. (2006), which combines a high robustness with a high statistical effi-
ciency. The auxiliary scale estimate s0 in (9) is given by s0 = medm+1≤t≤n |yt−yt|t−1|,
the median absolute deviation of the prediction errors. The loss function used in
(9) is
ρk(y) = min(k2, y2) = ψk(y)2, (10)
with k = 2.
3.2 Control chart
Similar as in subsection 2.2, we plot the one-step-ahead forecast errors et = yt−yt|t−1,
for t > n, in the control chart, but now with forecasts obtained from the robust Holt-
Winters algorithm. The control limits are determined from a robust scale estimate
of the one-step-ahead forecast errors from the training period. We use again the
τ -estimator of scale
τ 2 = cks20
n−m
n∑
t=m+1
ρk
(
et
sT
)
, (11)
where s0 = medm+1≤t≤n |et|, and ρk as in (10). The consistency factor ck in the
above equation ensures that τ 2 is a consistent estimator of the population variance
of a normal distribution. For k = 2, we have ck = 1.404. We construct approximate
6
(1 − α) control limits as
UCL = zα/2 × τ
LCL = − zα/2 × τ(12)
with zα/2 as the (1 − α/2)-quantile of the standard normal distribution. Hence,
the formula for the control limits (5) and (12) are similar, except that the robust
method uses a robust τ -scale estimator instead of a standard deviation (centered at
zero). The use of the τ -scale estimator guarantees the resistance to outliers in the
training period.
In the next section a simulation study is carried out to verify whether the con-
structed control limits for the standard and the robust Holt-Winters method have
comparable properties if the training sample is clean. In the presence of outliers in
the training sample, we expect the robust method to perform better.
4 Simulation study
In this simulation study we take a closer look at the size, the power and the false
detection rate of the proposed robust control chart, and compare them to the same
characteristics of the non-robust version. The size of a control chart is the proba-
bility that an observation in the test sample falls out of the control limits when the
process is actually in control, i.e. when no outliers are present in the test sample.
It is supposed to be approximately 5%, when taking α = 0.05 in (5) and (12). The
power is the probability that an alarm observation is detected. In a simulation study,
the power equals the proportion of generated outliers in the test sample that are
detected. The percentage of observations in the test sample that are not generated
as outliers, but that are falsely identified as alarm observations, is the false detection
rate. The false detection rate is like the size of the control charts, but for a process
that is not fully under control.
4.1 Simulation design
We generate series {yt} from a local linear trend model. This model generates
non-stationary and correlated time series, and allows for trends. It is a reasonable
model for many series encountered in the practice of business and economics. At
this model the forecast errors of the Holt-Winters method form an i.i.d. sequence
7
from a normal distribution (e.g. Chatfield and Yar (1988)). The local linear trend
model is defined as
Yt = αt + εt , εt ∼ N(0, σ2ε) , (13)
where αt, the local level, and βt, the local trend, are given by
{
αt = αt−1 + βt−1 + ηt, ηt ∼ N(0, σ2η),
βt = βt−1 + νt, νt ∼ N(0, σ2ν). ,
(14)
The noise terms εt, ηt and νt are independent, and serially uncorrelated. In the
simulation study, we take σε = 1 and ση = σν = 0.1, leading to a smaller variance
for the local level and trend compared to the variance in the measurement equation
(13).
The first n observations from a generated series constitute the training sample,
including a startup period of length m = 10. The next 200 generated observations
constitute the test sample. From the training data the control limits are constructed
as outlined in Sections 2 and 3. Outliers, hence alarm observations, are induced in
the test sample by adding the value kσε to 10% of randomly selected observations in
the test sample. For simulating the power, we take k = 5, corresponding to outliers
of moderate size. To study the robustness of the size and power with respect to
outliers in the training sample, the outliers are generated in the same way as for the
test sample. The percentage of outliers in the training data are taken as 0% (clean),
2%, and 5%, respectively.
4.2 Simulation Results
The simulated power and size are summarized in Table 1 for the control charts based
on the standard and the robust Holt-Winters algorithm. The reported numbers are
averages over 1000 simulation runs. Training samples of length n = 50 and n = 100
are considered. The simulation study, although rather modest in the number of
considered simulation settings, is quite time consuming, since the optimal selection
of the smoothing parameters, see (4) and (9), needs to be performed for every single
time series.
As we see from Table 1 there is a small size distortion in the clean case, when
no outliers are present in the training sample, and this for both methods. For the
larger sample size, the simulated size is already much closer to 5%. In the presence
of outliers in the training sample, the size of the non-robust chart is close to zero
in most cases, while the size of the robust method is kept much more stable. The
8
Table 1: Size and Power of the control chart, with α = 5%, for the standard (HW)
and the robust Holt-Winters (RHW) procedures, for different sizes of the training
sample (50 and 100), and for different percentages of outliers in training sample.
Size Clean 2% 5%
n HW RHW HW RHW HW RHW
50 .073 .086 .044 .073 .026 .067
100 .059 .063 .027 .052 .009 .037
Power Clean 2% 5%
n HW RHW HW RHW HW RHW
50 .903 .900 .843 .874 .784 .850
100 .903 .902 .842 .881 .757 .853
breakdown of the size for the non-robust method is due to two reasons. The main
reason is that the scale S in (5) is inflated due to the outliers, resulting in too
wide control limits. Furthermore, the outliers cause a bias in the selection of the
smoothing parameters, leading to a less optimal smoothing.
Consider now the power in Table 1. If the training sample is clean, the two meth-
ods have a comparable power to detect the generated alarm observations. When 2%
or 5% of outliers are present in the training sample, they do affect the detection
power of the control charts, and we see that the robust Holt-Winters clearly out-
performs the standard one. The difference in power is about 10% in presence of
only 5% of moderate outliers in the training sample. When increasing k, the size of
the outliers, and the percentage of outliers in the training sample, these differences
become even much more pronounced.
Next, we compare the false detection rate of the two control charts. We follow the
same simulation scheme as before, but only consider clean training data to ensure
that the size of the two methods remains comparable. We vary the magnitude of
the outliers by taking k = 5, 10, 15 and 20. A first observation is that the false
detection rate does not seem to depend much on n. Furthermore, for k = 5 the
robust method is only slightly better than the non-robust method. For larger values
of k, in contrast, the robust procedure has a much lower false detection rate. This
is due to the following reason. For large values of k, an outlying observation in the
test sample affects the outcome of the next forecast in the Holt-Winters procedure.
9
Table 2: False detection rate of the control chart, with α = 5%, for the standard
(HW) and the robust Holt-Winters (RHW) algorithm, and for different lengths n
of the training sample. The training data are clean, and outliers are created in the
test data by shifting 10% of the observations upwards by k units.
HW RHW
shift size k = 5 k = 10 k = 15 k = 20 k = 5 k = 10 k = 15 k = 20
n = 50 0.097 0.167 0.206 0.232 0.084 0.088 0.083 0.085
n = 100 0.093 0.161 0.203 0.229 0.079 0.080 0.081 0.082
Hence it is likely to induce a large forecasting error for the next observation as well,
even if the latter one is not an outlier. The propagation effect of outliers does not
occur with the robust approach, since it uses “cleaned observations” in the update
formulas (8).
5 Real data example
We illustrate the proposed method by means of a real data example. We use the
Housing Data from the book of Diebold (2001), containing monthly data for housing
starts and completions from January 1968 until June 1996. Figure 1 shows both
series and the corresponding control charts obtained by the standard and the robust
Holt-Winters procedure. The upper panel of every graph shows the observed data yt,
together with the smoothed series yt|t−1. The lower panel of every graph contains the
control chart. As explained in section 2, the prediction errors et are plotted in the
control chart. The vertical dashed lines in the chart indicate the end of the startup
period and of the training period. Control limits are indicated for the test period.
We used a startup period of length m = 20, a training period of length n = 150.
The other 192 observations serve as the test sample. We immediately notice two
large outliers, one near 1971 and another near 1977, in the training sample of the
housing starts. Moreover, both time series may contain other smaller outliers. By
applying robust exponential smoothing, we expect that the results remain stable in
the presence of these outliers.
We first take a closer look at the housing starts in Figure 1(a)-(b). The smoothing
parameters (λ1, λ2) selected from the training period are (0.3, 0.2) for the standard
10
(a)
0.5
1.5
2.5
Standard H−W
Hou
sing
sta
rts
1970 1975 1980 1985 1990 1995
−2.
0−
1.0
0.0
1.0
fore
cast
err
ors
(b)
0.5
1.5
2.5
Robust H−W
Hou
sing
sta
rts
1970 1975 1980 1985 1990 1995
−2.
0−
1.0
0.0
1.0
fore
cast
err
ors
(c)
0.5
1.5
2.5
Standard H−W
Hou
sing
com
plet
ions
1970 1975 1980 1985 1990 1995
−0.
40.
00.
20.
4
fore
cast
err
ors
(d)
0.5
1.5
2.5
Robust H−W
Hou
sing
com
plet
ions
1970 1975 1980 1985 1990 1995
−0.
40.
00.
20.
4
fore
cast
err
ors
Figure 1: In each graph, the upper panel presents the observed data (dots) and the smoothed
series (solid line), while the lower panel presents the control chart, with forecast errors (points)
and control limits (dashed lines). The housing starts series - containing outliers in the training
period - is analyzed with standard Holt-Winters in (a) and with robust Holt-Winters method in
(b). Similarly for the housing completions series. This series does not contain outliers in the
training period, explaining the resemblance between (c) and (d).
11
method and (0.4, 0.2) for the robust method. These parameters are reasonable,
since they are close to zero and the parameter corresponding to the trend is smaller
than the one for the level. The standard method tends to select smaller parameters
than the robust version, since outliers may bias the parameters towards zero. The
direct effect of the large outliers in the standard Holt-Winters case can be seen
at the time points subsequent to the large outliers. The forecasts for these time
points are attracted by the outlier and therefore, in this case, the values are too
small. The robust method clearly yields better forecasts in the outlier’s subsequent
time points. Also in terms of mean squared one-step-ahead forecast errors (MSFE),
computed over the test period, the robust method outperforms the standard one
with an MSFE of 0.0144 against 0.0175. The control chart shows that the control
limits are inflated by the outliers in the non-robust case, and all observations in
the test sample remain within the control bounds. This leads to a loss of power of
the non-robust method to detect outliers. For example, the alarm observation in
January 1984 is only detected by the robust method.
The housing completions in Figure 1(a)-(b), do not show important outliers in
the training period. The selected smoothing parameters for this series are (0.5, 0.2)
in the standard case and (0.6, 0.3) in the robust case. Similar to the housing starts,
the standard parameters are somewhat smaller than the robust ones. The MSFE for
both methods are comparable, with 6.01×10−3 in the standard case and 6.09×10−3
in the robust case. Since the series do not contain important outliers, both the
standard and the robust control charts look similar. Nevertheless, the robust control
chart has slightly tighter control limits. The percentage of observations exceeding
the control limits is 2% for the standard method, which seems overly optimistic, and
6.1% for the robust method, closer to the expected size of α = 5%.
We conclude that it is worthwhile to use the robust method. If there are no
major outliers, the performance of both methods is quite similar. However, if major
outliers are present, in particular in the training period, the robust method is more
reliable.
6 Conclusion
In this paper, we propose to use the one-step-ahead forecast errors of the Holt-
Winters forecasting method as the input of a control chart for monitoring non-
stationary processes. Holt-Winters forecasting is a frequently used method for uni-
variate time series with a stochastic trend and a seasonality pattern. In this study,
12
we did not explicitly deal with seasonality, but the method can easily be adapted for
that purpose. The main advantages are that the method is very easy to implement
and still performs well compared to other more complicated forecasting techniques.
We compared the control charts based on the standard Holt-Winters method with
a robust version.
The simulation study shows that with clean training data, the standard and
robust control charts perform almost equally well in terms of size and power. On the
other hand, when we add a small amount of moderate outliers to the training sample
the non robust method looses power, and suffers from severe size distortion. Even
when the training data are clean, the false detection rate of the robust procedures is
much lower if major outliers are present in the test sample. Finally, the application
to real data shows that the proposed robust method is easy to put in practice. The
graphical displays in Figure 1, representing both the observed series and the control
chart in the same graph, are an easy-to-use visual aid for managers to monitor
business processes.
There are several open questions, which we consider as areas for future research.
If a structural break appears in the time series, the control chart should be readjusted
and new control limits need to be computed. Another limitation of the presented
approach is the normality assumption on the forecast errors (excluding outliers).
Different distributional assumptions will lead to different formula for the control
limits. We refer to Yang and Chen (2009) and Chen and Yeh (2009) for recent
contributions to these problems, although in different settings.
References
Alfaro, J. and Ortega, J. (2009). A comparison of robust alternatives to Hotelling’s
T 2 control chart. Journal of Applied Statistics , 36(12), 1385–1396.
Alwan, L. C. and Roberts, H. V. (1988). Time-series modeling for statistical process
control. Journal of Business & Economic Statistics , 6(1), 87–95.
Box, G. and Ramırez, J. (1992). Cumulative score charts. Quality and Reliability
Engineering , 8(1), 17–27.
Chatfield, C. and Yar, M. (1988). Holt-Winters forecasting: some practical issues.
The Statistician, 37(2), 129–140.
Chen, F. L. and Yeh, C. H. (2009). Economic statistical design of non-uniform
13
sampling scheme X bar control charts under non-normality and Gamma shock
using genetic algorithm. Expert Systems with Applications , 36(5), 9488–9497.
De Gooijer, J. and Hyndman, R. (2006). 25 years of time series forecasting. Inter-
national Journal of Forecasting , 22(3), 443–473.
Diebold, F. X. (2001). Elements of forecasting . South-Western, second edition.
Gardner, E. (2006). Exponential smoothing: the state of the art - part ii. Interna-
tional Journal of Forecasting , 22(4), 637–666.
Gelper, S., Fried, R., and Croux, C. (2009). Robust forecasting with exponential
and Holt-Winters smoothing. Journal of Forecasting , 29, 285–300.
Jiang, W., Tsui, K.-L., and Woodall, W. H. (2000). A new SPC monitoring method:
the arma chart. Technometrics , 42(4), 399–410.
Maronna, R. A., Martin, R. D., and Yohai, V. J. (2006). Robust statistics: theory
and methods . Wiley.
Nembhard, H. B. and Changpetch, P. (2007). Directed monitoring using cuscore
charts for seasonal time series. Quality and Reliability Engineering International ,
23, 219–232.
Rocke, D. M. (1989). Robust control charts. Technometrics , 31(2), 173–184.
Rocke, D. M. (1992). Xq and rq charts: robust control charts. The Statistician,
41(1), 97–104.
Stefatos, G. and Hamza, A. B. (2009). Fault detection using robust multivariate
control chart. Expert Systems with Applications , 36(3), 5888–5894.
Tatum, L. G. (1997). Robust estimation of the process standard deviation for contol
charts. Technometrics , 39(2), 127–141.
Vander Wiel, S. (1996). Monitoring processen that wander using integrated moving
everage models. Technometrics , 38(2).
Vargas, J. A. (2003). Robust estimation in multivariate control charts for individual
observations. Journal of Quality Technology , 35(4).
Wang, F. (2005). A simple data transformation of auto-correlated data for SPC.
International Journal of Production Research, 43(5), 981–989.
14